Multi-round dialogue processing method and device for real-time news based on large model

Abstract

The present invention provides a method and device for processing multi-round dialogues for real-time news based on a large model, which relates to the technical field of multi-round dialogue processing for real-time news, and includes: based on a first preset template in a multimodal large model, determining the current summary information of the real-time news event corresponding to the current request information; filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into a second preset template of the multimodal large model, and determining the first response of each current summary information in the multi-round dialogue under the action of each historical request information, and the second response of each historical summary information in the multi-round dialogue under the action of each current request information; inputting the multimodal large model according to the first response, the second response and the historical response, determining the response to the current request information in the multi-round dialogue; so as to alleviate the technical problem of poor accuracy of the request answer of real-time news in the multi-round dialogue.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-round dialogue processing for real-time news, and in particular to a multi-round dialogue processing method and device for real-time news based on a large model. Background Art

[0002] With the emergence of multimodal large models (multimodal large language models), AI can communicate smoothly with people and has shown leaps and bounds in semantic understanding and logical reasoning.

[0003] Current multimodal large models generally obtain answers to user questions by pre-setting an updateable knowledge base; however, such answers have a lag and cannot meet the user's question requirements regarding real-time news events; in addition, if the user conducts multiple rounds of dialogue requests regarding real-time news events, the lag in the answers will be accumulated, thereby greatly reducing the accuracy of the responses to the requests in multiple rounds of dialogue. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for processing multi-round dialogues for real-time news based on a large model, so as to alleviate the technical problem of poor accuracy in answering requests for real-time news in multi-round dialogues.

[0005] In a first aspect, an embodiment of the present invention provides a multi-round dialogue processing method for real-time news based on a large model, comprising:

[0006] In response to current request information for a real-time news event, determining whether the current request information constitutes a multi-round dialogue with previous request information;

[0007] If not, controlling the multimodal large model to output a response to the current request information;

[0008] If yes, then determining the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model;

[0009] Filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multiple rounds of conversations into a second preset template of the multimodal large model;

[0010] Based on the second preset template of the multimodal large model, determining a first response of each current summary information in the multiple rounds of dialogue under the action of each historical request information, and a second response of each historical summary information in the multiple rounds of dialogue under the action of each current request information;

[0011] The multimodal big model is inputted according to the first response, the second response and the historical response to determine the response to the current request information in the multi-round dialogue.

[0012] In combination with the first aspect, an embodiment of the present invention provides a first possible implementation manner of the first aspect, wherein the step of filling the current summary information, current request information, historical summary information, historical request information, and historical responses in the multi-round dialogue into the second preset template of the multimodal large model includes:

[0013] Acquire each historical request information in the multiple rounds of dialogues, historical summary information of each real-time news event corresponding to each historical request information, and historical responses corresponding to each historical request information;

[0014] Configuring a field group in a second preset template of the multimodal macromodel according to a preset length and a preset content type;

[0015] Extracting target information from each historical request information, each historical summary information of each real-time news event corresponding to the historical request information and each historical response corresponding to the historical request information, and each current request information and each current summary information of each real-time news event corresponding to the current request information; wherein the target information meets the requirements of the preset content type and the preset length;

[0016] Each historical request information and each current summary information are automatically filled into the first field group, and each current request information and each historical summary information are automatically filled into the second field group.

[0017] In combination with the first aspect, an embodiment of the present invention provides a second possible implementation of the first aspect, wherein the step of extracting target information from each historical request information, each historical summary information of each real-time news event corresponding to the historical request information and each historical response corresponding to the historical request information, and each current request information and each current summary information of each real-time news event corresponding to the current request information, comprises:

[0018] Determine whether there is abnormal text from which target information cannot be extracted in each historical request information, each historical summary information of each real-time news event corresponding to the historical request information, each historical response corresponding to the historical request information, each current request information, and each current summary information of each real-time news event corresponding to the current request information;

[0019] If it does not exist, then execute the step of extracting target information;

[0020] If it exists, the abnormal text is displayed on a graphical user interface, and target information of the abnormal text is generated in combination with a user confirmation instruction input into the graphical user interface.

[0021] In combination with the first aspect, an embodiment of the present invention provides a third possible implementation manner of the first aspect, wherein the step of filling the current summary information, current request information, historical summary information, historical request information, and historical responses in the multi-round dialogue into the second preset template of the multimodal large model further includes:

[0022] In the process of filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multiple rounds of conversations into the second preset template of the multimodal large model, if the number of field groups of the second preset template reaches a preset number threshold, the preset number of field groups farthest from the current filling time will be deleted or merged.

[0023] In combination with the first aspect, an embodiment of the present invention provides a fourth possible implementation of the first aspect, wherein, in response to current request information for a real-time news event, the step of determining whether the current request information constitutes a multi-round dialogue of previous request information includes:

[0024] Determining whether the current request information and the previous request information constitute a multi-round dialogue by judging whether the time interval between inputting the current request information and inputting the previous request information is within a preset time range;

[0025] or,

[0026] By judging whether the matching degree of the current request information with the previous request information, the response and summary information corresponding to the previous request information respectively meets the preset matching degree threshold, it is determined whether the current request information and the previous request information constitute a multi-round dialogue.

[0027] In combination with the first aspect, the embodiment of the present invention provides a fifth possible implementation of the first aspect, wherein the step of controlling the multimodal large model to output a response to the current request information includes:

[0028] In response to current request information input to the graphical user interface, controlling the real-time search engine to retrieve at least one real-time news event corresponding to the current request information;

[0029] Extracting corresponding key information from each of the real-time news events, and inputting the key information into the multimodal big model, and determining summary information corresponding to each of the real-time news events; wherein the multimodal big model includes a first preset template for extracting the summary information, and the first preset template is determined according to the format requirements and attribute requirements of the summary information;

[0030] Determining the authenticity of the real-time news event corresponding to each of the summary information according to the verification rules and verification weights corresponding to each news element of the summary information in the third preset template;

[0031] Determining a target real-time news event corresponding to the request information from the real-time news events according to the authenticity of the real-time news event and the summary information;

[0032] The request information and summary information of the target real-time news event are input into the multimodal macro model, and a response to the current request information is output under the action of the second preset template.

[0033] In combination with the first aspect, an embodiment of the present invention provides a sixth possible implementation of the first aspect, wherein the step of determining the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model includes:

[0034] Determine, according to the preset format requirements and preset attribute requirements of the current summary information, a first preset template for each real-time news event corresponding to the current request information; wherein the attribute requirements include one or more of the following: news elements of the real-time news event required to be displayed, the number of words of the summary required to be displayed, and the news type of the real-time news event required to be displayed;

[0035] Extracting corresponding key information from each of the real-time news events and inputting it into a multimodal large model;

[0036] Extracting corresponding first key information from the key information of each real-time news event according to the attribute requirements;

[0037] Each of the first key information is processed according to the format requirements to obtain summary information of each of the real-time news events; wherein the format requirements include a subject, a predicate, an object and a qualifier, and the qualifier is located before the subject and the object.

[0038] In combination with the first aspect, an embodiment of the present invention provides a seventh possible implementation manner of the first aspect, wherein the step of determining, based on the second preset template of the multimodal large model, a first response of each current summary information in the multi-round dialogue under the action of each historical request information, and a second response of each historical summary information in the multi-round dialogue under the action of each current request information includes:

[0039] Traversing the current summary information respectively corresponding to each historical request information in the first field group, and generating a first response for each current summary information under the action of the historical request information;

[0040] The historical summary information respectively corresponding to each current request information in the second field group is traversed to generate a second response of each historical summary information under the effect of the current request information.

[0041] In combination with the first aspect, an embodiment of the present invention provides an eighth possible implementation manner of the first aspect, wherein the step of inputting the multimodal big model according to the first response, the second response, and the historical response to determine the response to the current request information in the multi-round dialogue includes:

[0042] Based on the relevance of the current request information with the first response, the second response and the historical response, respectively assigning a weight ratio to the first response, the second response and the historical response;

[0043] According to the matching degrees and the weight ratios of the first response, the second response and the historical response with the current request information respectively, the responses to the current request information in the multiple rounds of dialogue are determined and output.

[0044] In a second aspect, an embodiment of the present invention further provides a multi-round dialogue processing device for real-time news based on a large model, comprising:

[0045] A judgment module, in response to current request information for real-time news events, judges whether the current request information constitutes a multi-round dialogue with the previous request information;

[0046] The first response module controls the multimodal large model to output a response to the current request information if no;

[0047] The first determination module, if yes, determines the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model;

[0048] A filling module fills the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into a second preset template of the multimodal large model;

[0049] A second determination module determines, based on a second preset template of the multimodal large model, a first response of each current summary information in the multiple rounds of dialogue under the action of each historical request information, and a second response of each historical summary information in the multiple rounds of dialogue under the action of each current request information;

[0050] The second response module inputs the multimodal large model according to the first response, the second response and the historical response, and determines the response to the current request information in the multi-round dialogue.

[0051] The embodiment of the present invention provides a multi-round dialogue processing method and device for real-time news based on a large model. First, it is determined whether the current request information for the real-time news event belongs to the first request of the current multi-round dialogue, or belongs to other requests in the current multi-round dialogue that exist before the current request information; if it is the first request, a real-time request information response is determined according to the corresponding real-time news event; if it is not the first request, that is, the current request information and the previous request information belong to the same multi-round dialogue, the current summary information of each real-time news event corresponding to the current request information is extracted using a first preset template, and then the current request, current summary, historical request, historical summary and historical response in the current multi-round dialogue are filled into a second preset template to obtain the first response of each current summary to the historical request information, the second response of each historical summary to the current request information, and the historical response of each historical summary information to the historical request information; based on the response set obtained by crossing these summary information with related relationships in the same multi-round dialogue, a response result with higher accuracy for the current request information can be obtained.

[0052] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A flow chart of a multi-round dialogue processing method for real-time news based on a large model provided by an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of functional modules of a multi-round dialogue processing device for real-time news based on a large model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] At present, multimodal large models generally ignore the real-time nature of responses to real-time news events, which leads to lags and errors in the response results. For example, the number of gold medals won by Country X in the Olympics is real-time and the number of gold medals changes during the Olympic Games.

[0059] However, if a large multimodal model processes multiple rounds of conversations on real-time news events, such errors and lags will be superimposed, further leading to lower accuracy of the response results.

[0060] Based on this, an embodiment of the present invention provides a method and device for processing multi-round dialogues for real-time news based on a large model, which can determine the response to the current request that has a dependency relationship with the multi-round dialogue through the current request and historical summary, and the cross-answer between the historical request and the current summary.

[0061] To facilitate understanding of this embodiment, a multi-round dialogue processing method for real-time news based on a large model disclosed in an embodiment of the present invention is first introduced in detail. This method can be applied to intelligent devices such as controllers, host computers, servers, and touch-screen interactive terminals.

[0062] Figure 1 A flow chart of a multi-round dialogue processing method for real-time news based on a large model provided by an embodiment of the present invention;

[0063] Reference Figure 1 , the method mainly comprises the following steps:

[0064] Step S102, in response to the current request information for the real-time news event, determining whether the current request information constitutes multiple rounds of dialogue with the previous request information.

[0065] The current request information can be input through a graphical user interface, which can be understood as an interface integrated in the above-mentioned smart device for facilitating user interaction; the request information in the embodiment of the present invention can be understood as a request statement for a real-time news event; the request statement can include multiple modal forms, such as voice, text, video, etc.; after each modal request statement is converted into text, the statement text includes first text information for locating the real-time news event, and second text information for characterizing a corresponding request for the text information, so that at least one real-time news event can be retrieved and captured based on the request information;

[0066] It should be noted that, based on the association relationship between the current request information and the previous request information, it is determined whether the two are in the same dialogue, that is, constitute a multi-round dialogue; as an optional embodiment, based on the association relationship between the current request information and the response to the previous request information, it is determined whether the current request information and the previous request information constitute a multi-round dialogue.

[0067] Step S104: if not, control the multimodal large model to output a response to the current request information.

[0068] Here, if the current request information and the previous request information belong to different conversations, that is, the current request information is the first question in the current conversation, a real-time response is made based on the real-time news event targeted by the request information.

[0069] Step S106: If yes, then based on the first preset template in the multimodal large model, determine the current summary information of the real-time news event corresponding to the current request information.

[0070] Here, if the current request information and the previous request information belong to the same conversation, that is, the current request information and the previous request information constitute multiple rounds of conversations, it is necessary to determine the dependency between the current request information and the previous request information and the response to the previous request information, and then obtain an accurate response to the request information that was not raised for the first time in the multiple rounds of conversations; first, obtain at least one current summary information corresponding to the request information through a first preset template, and each current summary information corresponds to a real-time news event.

[0071] Step S108, filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into the second preset template of the multimodal large model.

[0072] It should be noted that, through the steps of the above embodiment, it is possible to know whether the current request information and the previous request information belong to the same multi-round dialogue. In this way, it is possible to know whether the previous request information is associated with its corresponding previous request information, and so on to obtain all the historical request information in the current multi-round dialogue, and the corresponding historical response of each historical request information;

[0073] Among them, the second preset template (after information filling) guides the multimodal large model to respond to the request information according to the pre-configured processing response logic corresponding to the various request types; for example, the third preset template determines the request type of the request information by identifying specific keywords in the request information, and selects the corresponding reply rule according to the corresponding request type. Under the action of the reply rule, an answer to the request information is generated based on the summary information of the target real-time news event.

[0074] Here, the second preset template itself also presets configuration filling rules so that the current summary information, current request information, historical summary information, historical request information and historical response information are filled in the second preset template in an agreed manner according to the specified filling requirements.

[0075] Step S110, based on the second preset template of the multimodal large model, determine the first response of each current summary information in the multi-round dialogue under the action of each historical request information, and the second response of each historical summary information in the multi-round dialogue under the action of each current request information.

[0076] According to the second preset template filled in an agreed manner, it is possible to make a first response to the historical request information using each current summary information, and to make a second response to the current request information using each historical summary information.

[0077] Step S112, inputting the multimodal large model according to the first response, the second response and the historical responses, and determining the response to the current request information in the multi-round dialogue.

[0078] It should be noted that the multimodal large model, that is, the multimodal large language model has the function of generating, answering and understanding natural language, and the natural language is not limited to text form, but can also include various modal forms, such as images, videos, audio, etc.

[0079] Under the action of the second preset template, the multimodal large model in the embodiment of the present invention can reply to the current request based on the historical summary, reply to the historical request based on the current summary, and reply to the historical request based on the historical summary. It can obtain richer response corpus in multi-round conversations with dependencies, and reply to the current request based on these response corpora, which can reflect the correlation between the response to the current request information and the multi-round conversations, thereby ensuring the accuracy of the response to the current request information in the multi-round conversations.

[0080] In a preferred embodiment of actual application, it is first determined whether the current request information for the real-time news event belongs to the first request of the current multi-round dialogue, or to other requests in the current multi-round dialogue that existed before the current request information; if it is the first request, a real-time request information response is determined based on the corresponding real-time news event; if it is not the first request, that is, the current request information and the previous request information belong to the same multi-round dialogue, the current summary information of each real-time news event corresponding to the current request information is extracted using the first preset template, and then the current request, current summary, historical request, historical summary and historical response in the current multi-round dialogue are filled into the second preset template to obtain the first response of each current summary to the historical request information, the second response of each historical summary to the current request information, and the historical response of each historical summary information to the historical request information; based on the response set obtained by crossing these summary information with related relationships in the same multi-round dialogue, the response to the current request that has a dependency relationship with the multi-round dialogue is determined, so that a response result with higher accuracy for the current request information can be obtained.

[0081] In some embodiments, the determination in step S102 whether the current request information constitutes a multi-round dialogue with the previous request information may be implemented in the following manners, including:

[0082] Step 1.1), by judging whether the time interval between inputting the current request information and inputting the previous request information is within a preset time range, it is determined whether the current request information and the previous request information constitute a multi-round dialogue.

[0083] Here, if the time interval between the time when the current request information is input into the graphical user interface and the time interval between the time when the previous request information is input into the graphical user interface meets the preset time range, such as 1 minute, 3 minutes, 5 minutes, etc., the current request information and the previous request information can be identified as the same multi-round dialogue.

[0084] or,

[0085] Step 1.2), by judging whether the matching degree of the current request information with the previous request information, the response corresponding to the previous request information and the summary information respectively meets the preset matching degree threshold, it is determined whether the current request information and the previous request information constitute a multi-round dialogue.

[0086] Here, as an optional embodiment, based on the comparison of the matching degree of the current request information and the previous request information, the summary information of each real-time news event corresponding to the previous request information, and the corresponding response, and the preset matching degree threshold, it can be determined that the current request information and the previous request information that meet the matching degree requirements constitute the same multi-round dialogue.

[0087] As another optional embodiment, it is necessary to satisfy step 1.1) and step 1.2) at the same time, that is, when the time interval between the current request information and the previous request information meets the requirement, the current request information and the corresponding summary and response have a matching degree with the previous request information that meets the requirement. At this time, the current request information and the previous request information are identified as the same multi-round dialogue.

[0088] The method for determining whether the current request information and the previous request information constitute a multi-round dialogue may be selected based on the actual application situation and accuracy requirements.

[0089] It should be noted that the current request information may correspond to at least one real-time news event, and the real-time news event corresponding to the current request information can be determined through a real-time search engine, and the real-time news event corresponding to the previous request information is similar; in actual applications, it is also possible to determine whether the current request information and the previous request information belong to the same multi-round dialogue through the matching degree between the real-time news event 1 corresponding to the previous request information and the real-time news event 2 corresponding to the current request information;

[0090] It can be understood that the operations such as the matching degree determination in the aforementioned steps in the embodiments of the present invention can be implemented based on the matching model in the neural network.

[0091] In some embodiments, if the current request information and the previous request information do not belong to the same multi-round dialogue, the current request information belongs to the first request of the current multi-round dialogue. At this time, step S104 can be determined based on the precise real-time nature of the real-time news event to ensure the reliability of the response result, specifically including:

[0092] Step 2.1), in response to current request information input into the graphical user interface, control the real-time search engine to retrieve at least one real-time news event corresponding to the current request information.

[0093] In step 2.2), the corresponding key information is extracted from each real-time news event, and the key information is input into the multimodal large model to determine the summary information corresponding to each real-time news event.

[0094] Among them, the multimodal large model includes a first preset template for extracting summary information. The first preset template of the multimodal large model converts the summary information into text form and performs text extraction, wherein the format requirements and attribute requirements satisfied by the expected summary information are pre-set; the format requirements can be understood as the arrangement of the text in the summary information, and the attribute requirements can be understood as the content information of the text in the summary information, and the attributes corresponding to the content information; and the first preset template is determined according to the format requirements and attribute requirements of the summary information; the first preset template is used to guide the multimodal large model to further extract summary information from the key information of each real-time news event; here, the key information can be understood as the information that is initially captured from each real-time news event and can reflect the main content of the real-time news event, usually including at least one complete news paragraph in the real-time news event.

[0095] Step 2.3), according to the verification rules and verification weights corresponding to each news element of the summary information in the third preset template, the authenticity of the real-time news event corresponding to each summary information is determined.

[0096] It should be noted that the current real-time news events are captured only because they are related to the request information, and the authenticity of such real-time news events themselves cannot be guaranteed. In this embodiment of the present invention, the news elements corresponding to each real-time news event are verified by the third preset template to determine the authenticity of each real-time news event.

[0097] Among them, each news event includes at most six news elements, such as time, people, place, cause, process, and result; the third preset template is pre-configured with verification rules for each news element, as well as the verification weight corresponding to each news element, that is, whether each news element is verified or not plays a different role in the authenticity of the real-time news event.

[0098] Step 2.4), according to the authenticity and summary information of the real-time news event, determine the target real-time news event corresponding to the requested information from the real-time news events.

[0099] It can be understood that, by combining the verification of the authenticity of each real-time news event in the aforementioned embodiment steps and the summary information extracted from each real-time news event, the target real-time news event with the highest match to the requested information is determined; in other words, the target real-time news event is generally a real-time news event that meets the requested information and has a higher degree of authenticity.

[0100] In step 2.5), the request information and summary information of the target real-time news event are input into the multimodal large model, and a response to the current request information is output under the action of the second preset template.

[0101] Here, the multimodal large model also includes a second preset template, which can respond to the input request information for the target real-time news event and output reply information of the request information.

[0102] In other embodiments, if the current request information and the previous request information belong to the same multi-round dialogue, the current request information, the current summary and the historical requests and historical summaries in the multi-round dialogue are used to determine the accurate response of the current request information in the multi-round dialogue; first, as described in step S106, the current summary information is extracted based on the real-time news event corresponding to the current request information, specifically including:

[0103] Step 3.1), according to the preset format requirements and preset attribute requirements of the current summary information, determine the first preset template corresponding to each real-time news event of the current request information.

[0104] Here, the first preset template for each real-time news event can be predetermined. The first preset template corresponding to each real-time news event may be the same or different. The first preset template is determined according to the format requirements and attribute requirements of the summary information. In other words, the expected requirements for the summary information corresponding to each real-time news event may be different, resulting in different real-time news events having different first preset templates.

[0105] As an optional embodiment, a preset initial template corresponding to each real-time news event can be determined based on the source URL and news type of each real-time news event; each preset initial template is then processed based on the required display format and required display attributes of the summary information to obtain a first preset template corresponding to each real-time news event.

[0106] In step 3.2), the corresponding key information is extracted from each real-time news event and input into the multimodal large model.

[0107] Here, the original disclosure of each real-time news event may contain some content that is not related to the news event itself, which is mainly reflected in the news event display paragraphs on the website page; for example, the main content of the news is presented in the second and third paragraphs of the page, and the first paragraph of the page is non-critical information that can be deleted; the embodiment of the present invention extracts the corresponding key paragraphs related to the main body of the news event from each real-time news event based on the request information, forms key information, and then inputs it into the multimodal large model.

[0108] Step 3.3), extract the corresponding first key information from the key information of each real-time news event according to the attribute requirements.

[0109] The attribute requirements include one or more of the following: news elements of real-time news events that require to be displayed, the number of words of abstracts that require to be displayed, and the news types of real-time news events that require to be displayed;

[0110] Here, the key paragraphs including at least one news event body are extracted in the above steps, and several news elements corresponding to the real-time news event of a specific news type are extracted according to the attribute requirements, and the first key information is presented according to a specific number of words in the summary;

[0111] For example, news types include sports, games, humanities, entertainment, society and the like. In the aforementioned embodiment, each real-time news event that matches the request information may correspond to a variety of different news types, or each real-time news event may include multiple news types. Here, the target news type of each real-time news event can be determined according to the attribute requirements in the first preset template, and the first key information only extracts information from key paragraphs of the real-time news event of the target news type.

[0112] In step 3.4), each first key information is processed according to the format requirements to obtain summary information of each real-time news event.

[0113] Among them, the format requirements include subject, predicate, object and qualifier, and the qualifier is placed before the subject and object.

[0114] Here, on the basis that the first key information meets the attribute requirements, the specific sentence components in the first key information are arranged according to the format requirements; for example, the first key information "August 5 is the tenth competition day of the A Olympic Games, and 19 Olympic gold medals will be produced on this day." After processing, the summary information "The tenth competition day of the A Olympic Games - August 5 - will produce - 19 - Olympic gold medals" is obtained.

[0115] In some embodiments, the current summary information corresponding to the current request information and the historical summary information corresponding to the historical request information are filled in accordance with the constraint conditions preset in the second preset template. The step S108 can also be implemented by the following steps, including:

[0116] Step 4.1) obtains each historical request information in multiple rounds of dialogue, the historical summary information of each real-time news event corresponding to each historical request information, and the historical response corresponding to each historical request information.

[0117] According to step S102 in the above embodiment, it is possible to know whether each request information input at the current moment and the request information input at the previous moment belong to the same multi-round dialogue, so as to know all historical requests of each multi-round dialogue in this way;

[0118] For example, a multi-round dialogue includes two historical requests A and B. The current request input can be known to belong to the same multi-round dialogue as the previous request B. At this time, the previous request B is used as the current request to determine whether the current request B and its previous request A belong to the same multi-round dialogue. If they do, the historical request A is used as the current request to determine that it and the previous request input by the graphical user interface belong to different multi-round dialogues, that is, the historical request A is the first request information of the current multi-round dialogue, and the corresponding summary information and response are obtained according to each request information in the current multi-round dialogue.

[0119] In step 4.2), the field group in the second preset template of the multimodal large model is configured according to the preset length and the preset content type.

[0120] Here, the second preset template can be configured according to preset business requirements and scenarios, so that after configuration, only information that meets the preset content type and the preset length can be filled into the field group of the second preset template; the preset length may include 10 characters, 5 characters, etc., and the preset content type may include subject, predicate, object, qualifier and other content types.

[0121] Step 4.3), extract the target information from each historical request information, the historical summary information of each real-time news event corresponding to each historical request information and the historical response corresponding to each historical request information, as well as each current request information and the current summary information of each real-time news event corresponding to each current request information.

[0122] The target information meets the requirements of the preset content type and the preset length. It is understandable that the historical request information, historical summary information, historical responses, current request information and each current request information and current summary information may contain parts that do not meet the requirements of the second preset template, and the target information that meets the requirements is extracted and then filled into the second preset template.

[0123] As an optional embodiment, in order to ensure the filling reliability of the second preset template, in addition to the automatic filling method according to the preset length requirement and the preset content type and other constraints shown in step 4.3), the content to be filled can also be verified in step 4.4) to ensure the reliability of the verified content to be filled, which specifically includes:

[0124] Step 4.4.1), determine whether there is abnormal text that cannot extract the target information in each historical request information, the historical summary information of each real-time news event corresponding to each historical request information, the historical response corresponding to each historical request information, each current request information, and the current summary information of each real-time news event corresponding to each current request information.

[0125] Here, abnormal text can be understood as text information from which the target information cannot be extracted from historical request information, historical summary information, historical responses, current request information, and current summary information. For example, if the above information does not meet the preset content type or cannot meet the preset length requirements, and thus the target information cannot be generated, then such text information is considered abnormal text.

[0126] Step 4.4.2), if it does not exist, execute step 4.3) to extract the target information.

[0127] Here, if the historical request information, the historical summary information, the historical response, the current request information and the current summary information can all be used to extract the target information, they can be directly extracted respectively.

[0128] Step 4.4.3), if it exists, the abnormal text is displayed on the graphical user interface, and the target information of the abnormal text is generated in combination with the user confirmation instruction input into the graphical user interface.

[0129] Here, through the dialogue mechanism, the abnormal text can be popped up and displayed on the graphical user interface in the form of a dialog box, so that the user can input the corresponding control instruction information, and generate the target information of this abnormal text based on the control instruction information; for example, if the control instruction information is a confirmation instruction, the abnormal text is used as the target information, or if the control instruction information is an adjustment instruction, the target information is manually selected from the abnormal text according to the user's adjustment filter.

[0130] In step 4.5), each historical request information and each current summary information are automatically filled into the first field group, and each current request information and each historical summary information are automatically filled into the second field group.

[0131] Here, it can be understood that the target information of each historical request information and the target information of each current summary information are automatically filled into the first field group; each first field group includes the target information of a historical request information and the target information of a current summary information corresponding to the historical request information; and the historical request information may include multiple, and the current summary information may also include multiple, that is, each historical request information may also correspond to multiple current summary information, so it can be known that the first field group includes at least one;

[0132] Similarly, the target information of each current request information and the target information of each historical summary information are automatically filled into the second field group, and the second field group also includes at least one.

[0133] It should be noted that the second preset template pre-sets the cross-correspondence between the current request and the historical summary, and the cross-correspondence between the historical request and the current summary, and the same correspondence is stored and filled in the same field group. Among them, the historical response is stored in the third field group.

[0134] As an optional embodiment, in order to ensure the filling reliability of the second preset template, in addition to the manner of automatically filling the first field group and the second field group respectively shown in step 4.5), the number of the first field group and the second field group may be constrained by another constraint condition pre-set in the second preset template, including:

[0135] In step 4.6, in the process of filling the current summary information, current request information, historical summary information, historical request information and historical responses in multiple rounds of conversations into the second preset template of the multimodal large model, if the number of field groups of the second preset template reaches the preset number threshold, the preset number of field groups farthest from the current filling time will be deleted or merged.

[0136] It should be noted that the second preset template may not only include the field group information corresponding to the current multi-round dialogue, and in order to ensure the storage space of the second preset template, the number of field groups carried in the second preset template is configured in advance;

[0137] As an optional example, if the filling of the first field group and the second field group corresponding to the current multi-round conversations causes the number of field groups in the second preset template to reach the upper limit, the historical summary information that reaches the forgetting condition can be deleted, that is, a specific number of field groups that are farthest away from the current filling time can be deleted. It can be understood that the field group that is farthest away from the current filling time belongs to other multi-round conversations.

[0138] As another optional example, if the filling of the first field group and the second field group corresponding to the current multiple rounds of conversations causes the number of field groups in the second preset template to reach an upper limit, the historical summary information that reaches the forgetting condition, that is, the summary information of a specific number of field groups that are farthest away from the current filling time, can be fused; for example, the second preset template can be configured with fixed field groups and fused field groups, and the fixed fields in the summary information of the specific number of field groups that are farthest away from the current filling time are extracted and filled in the fixed field group, while the non-fixed fields are integrated and filled in the fused field group, so as to save the number of first field groups and second field groups.

[0139] In practical applications, a cross-type answer is generated according to the first field group and the second field group filled in the second preset template, respectively; as described in step S110, it includes:

[0140] Step 5.1), traverse the current summary information corresponding to each historical request information in the first field group, and generate a first response for each current summary information under the action of the historical request information.

[0141] Among them, the historical request information may include n pieces, and the current summary information also includes m pieces. Therefore, m*n first field groups will be generated in the second preset template, and m*n first field groups will be traversed to generate m*n first responses to each historical request information based on each current summary information.

[0142] Step 5.2) traverses the historical summary information corresponding to each current request information in the second field group, and generates a second response for each historical summary information under the effect of the current request information.

[0143] Similar to step 5.1), by traversing the second field group, multiple second responses to the current request information based on the historical summary information can be obtained.

[0144] For example, historical request information: how many gold medals did Country X win in the Olympics; historical summary 1: August 18, Xiao Li - diving athlete - won the 19th gold medal in the Olympics; historical summary 2: August 18 - 19 gold medals in the Olympics; current request information: which event did the latest one win; current summary: August 18 - Country X diving - won a gold medal;

[0145] The first response made by the first field group to the historical request information based on the current summary is: On August 18, Country X won a gold medal in diving; The second response made by the second field group to the current request information based on the historical summary is: a. On August 18, diving won the 19th Olympic gold medal; b. On August 18, 19 gold medals were won in the Olympic Games; Historical response: On August 18, Country X won 19 gold medals in the Olympic Games;

[0146] Based on the above embodiment, step S112 obtains a response to the current request information based on the first response, the second response and the historical response. The response is dependent on the historical requests, historical responses and historical summaries in the multi-round dialogue and has higher accuracy, including:

[0147] Step 6.1), based on the relevance of the current request information with the first response, the second response and the historical response, a weight ratio is assigned to the first response, the second response and the historical response respectively.

[0148] Here, the second answer based on the current request information has the highest correlation with the current request information, the first answer has the second highest correlation with the current request information, and the historical answer has the lowest correlation with the current request information. According to this correlation ranking, corresponding preset weight ratios are allocated, such as 45%, 35% and 20%.

[0149] Step 6.2) According to the matching degree and weight ratio of the first response, the second response and the historical response to the current request information, the response to the current request information in the multi-round dialogue is determined and output.

[0150] Here, according to the above example, it can be understood that the multimodal large model can determine the field / semantic matching degree of each response with the current request information based on the above response set, so that it can be known that the latest gold medal was produced on August 18, and the latest gold medal is the 19th one of Country X, produced in the diving event, and won by diving athlete Xiao Li; according to the current request information, the response with the highest matching degree can be obtained as the first response and the second response a, and then combined with the product of the weight proportions corresponding to each response, the response to the current request information "Which event was the latest one won in" in multiple rounds of dialogue can be obtained as "On August 18, diving won the 19th Olympic gold medal."

[0151] The embodiment of the present invention can obtain multiple groups of response sets with dependent associations in multiple rounds of dialogues based on the cross-answers of current request information and historical summary information, and historical request information and current summary information, and then based on the multimodal large model of the response set, the accurate response to the current request information can be obtained.

[0152] In some embodiments, Figure 2 As shown, an embodiment of the present invention provides a multi-round dialogue processing device for real-time news based on a large model, comprising:

[0153] A judgment module, in response to current request information for real-time news events, judges whether the current request information constitutes a multi-round dialogue with the previous request information;

[0154] The first response module controls the multimodal large model to output a response to the current request information if no;

[0155] The first determination module, if yes, determines the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model;

[0156] A filling module fills the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into a second preset template of the multimodal large model;

[0157] A second determination module determines, based on a second preset template of the multimodal large model, a first response of each current summary information in the multiple rounds of dialogue under the action of each historical request information, and a second response of each historical summary information in the multiple rounds of dialogue under the action of each current request information;

[0158] The second response module inputs the multimodal large model according to the first response, the second response and the historical response, and determines the response to the current request information in the multi-round dialogue.

[0159] Furthermore, the filling module is also used to obtain each historical request information in the multiple rounds of conversations, the historical summary information of each real-time news event corresponding to each historical request information, and the historical response corresponding to each historical request information; configure the field group in the second preset template of the multimodal large model according to the preset length and preset content type; extract target information from each historical request information, the historical summary information of each real-time news event corresponding to each historical request information, and the historical response corresponding to each historical request information, as well as each current request information and the current summary information of each real-time news event corresponding to each current request information; wherein the target information meets the requirements of the preset content type and the preset length; each historical request information is automatically filled in the first field group with each current summary information, and each current request information is automatically filled in the second field group with each historical summary information.

[0160] Furthermore, the filling module is also used to determine whether there is abnormal text from which the target information cannot be extracted in each historical request information, the historical summary information of each real-time news event corresponding to each historical request information, the historical response corresponding to each historical request information, each current request information, and the current summary information of each real-time news event corresponding to each current request information; if not, executing the step of extracting the target information; if so, displaying the abnormal text on a graphical user interface, and generating the target information of the abnormal text in combination with a user confirmation instruction input into the graphical user interface.

[0161] Furthermore, the filling module is also used to, in the process of filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multiple rounds of conversations into the second preset template of the multimodal large model, if the number of field groups of the second preset template reaches a preset number threshold, then the preset number of field groups farthest from the current filling time will be deleted or merged.

[0162] Furthermore, the judgment module is also used to determine whether the current request information and the previous request information constitute a multi-round dialogue by judging whether the time interval between inputting the current request information and inputting the previous request information is within a preset time range; or to determine whether the current request information and the previous request information constitute a multi-round dialogue by judging whether the matching degree of the current request information with the previous request information, the response and the summary information corresponding to the previous request information respectively meets a preset matching degree threshold.

[0163] Furthermore, the first response module is also used to control the real-time search engine to retrieve at least one real-time news event corresponding to the current request information input into the graphical user interface in response to the current request information; extract corresponding key information from each of the real-time news events, and input the key information into the multimodal big model to determine the summary information corresponding to each of the real-time news events; wherein the multimodal big model includes a first preset template for extracting the summary information, and the first preset template is determined according to the format requirements and attribute requirements of the summary information; according to the verification rules and verification weights corresponding to each news element of the summary information in the third preset template, the authenticity of the real-time news event corresponding to each of the summary information is determined; according to the authenticity of the real-time news event and the summary information, the target real-time news event corresponding to the request information is determined from the real-time news event; the request information and the summary information of the target real-time news event are input into the multimodal big model, and the response to the current request information is output under the action of the second preset template.

[0164] Furthermore, the first determination module is also used to determine the first preset template of each real-time news event corresponding to the current request information according to the preset format requirements and preset attribute requirements of the current summary information; wherein the attribute requirements include one or more of the following: the news elements of the real-time news event required to be displayed, the number of words of the summary required to be displayed, and the news type of the real-time news event required to be displayed; extracting corresponding key information from each real-time news event and inputting it into the multimodal large model; extracting corresponding first key information from the key information of each real-time news event according to the attribute requirements; processing each first key information according to the format requirements to obtain summary information of each real-time news event; wherein the format requirements include subject, predicate, object and qualifier, and the qualifier is located before the subject and the object.

[0165] Furthermore, the second determination module is also used to traverse the current summary information corresponding to each historical request information in the first field group, and generate a first response for each current summary information under the action of the historical request information; traverse the historical summary information corresponding to each current request information in the second field group, and generate a second response for each historical summary information under the action of the current request information.

[0166] Furthermore, the second response module is also used to assign weight ratios to the first response, the second response and the historical response based on the correlation between the current request information and the first response, the second response and the historical response; and determine and output responses to the current request information in the multiple rounds of conversations based on the matching degree of the first response, the second response and the historical response with the current request information and the weight ratio.

[0167] It can be understood that the specific operation methods of each functional module in this embodiment can refer to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0168] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0169] In addition, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0170] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0171] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those of ordinary skill in the art should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed by the present invention, or can easily conceive of changes, or make equivalent replacements for some of the technical features therein. Such modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention.

Claims

1. A multi-round dialogue processing method for real-time news based on a large model, characterized in that: include: In response to current request information for a real-time news event, determining whether the current request information constitutes a multi-round dialogue with previous request information; If not, controlling the multimodal large model to output a response to the current request information; If yes, then determining the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model; Filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multiple rounds of conversations into a second preset template of the multimodal large model; Based on the second preset template of the multimodal large model, determining a first response of each current summary information in the multiple rounds of dialogue under the action of each historical request information, and a second response of each historical summary information in the multiple rounds of dialogue under the action of each current request information; Inputting the multimodal big model according to the first response, the second response and the historical response, determining a response to the current request information in the multi-round dialogue; The step of filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into the second preset template of the multimodal large model comprises: Acquire each historical request information in the multiple rounds of dialogues, historical summary information of each real-time news event corresponding to each historical request information, and historical responses corresponding to each historical request information; According to a preset length and a preset content type, a field group in a second preset template of the multimodal large model is configured; the field group includes a first field group and a second field group; Extracting target information from each historical request information, each historical summary information of each real-time news event corresponding to the historical request information and each historical response corresponding to the historical request information, and each current request information and each current summary information of each real-time news event corresponding to the current request information; wherein the target information meets the requirements of the preset content type and the preset length; Automatically fill the target information of each historical request information and the target information of each current summary information into the first field group, and automatically fill the target information of each current request information and the target information of each historical summary information into the second field group; The step of determining, based on the second preset template of the multimodal large model, a first response of each current summary information in the multi-round dialogue under the action of each historical request information, and a second response of each historical summary information in the multi-round dialogue under the action of each current request information, comprises: Traversing the current summary information respectively corresponding to each historical request information in the first field group, and generating a first response for each current summary information under the action of the historical request information; The historical summary information respectively corresponding to each current request information in the second field group is traversed to generate a second response of each historical summary information under the effect of the current request information.

2. The method according to claim 1, characterized in that The step of extracting target information from each historical request information, each historical summary information of each real-time news event corresponding to the historical request information and each historical response corresponding to the historical request information, and each current request information and each current summary information of each real-time news event corresponding to the current request information comprises: Determine whether there is abnormal text from which target information cannot be extracted in each historical request information, each historical summary information of each real-time news event corresponding to the historical request information, each historical response corresponding to the historical request information, each current request information, and each current summary information of each real-time news event corresponding to the current request information; If it does not exist, then execute the step of extracting target information; If it exists, the abnormal text is displayed on a graphical user interface, and target information of the abnormal text is generated in combination with a user confirmation instruction input into the graphical user interface.

3. The method according to claim 1, characterized in that The step of filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into the second preset template of the multimodal large model also includes: In the process of filling the current summary information, current request information, historical summary information, historical request information and historical responses in the multiple rounds of conversations into the second preset template of the multimodal large model, if the number of field groups of the second preset template reaches a preset number threshold, the preset number of field groups farthest from the current filling time will be deleted or merged.

4. The method according to claim 1, characterized in that: In response to current request information for a real-time news event, the step of determining whether the current request information is a previous request information constituting a multi-round dialogue includes: Determining whether the current request information and the previous request information constitute a multi-round dialogue by judging whether the time interval between inputting the current request information and inputting the previous request information is within a preset time range; or, By judging whether the matching degree of the current request information with the previous request information, the response and summary information corresponding to the previous request information respectively meets the preset matching degree threshold, it is determined whether the current request information and the previous request information constitute a multi-round dialogue.

5. The method according to claim 1, characterized in that The step of controlling the multimodal large model to output a response to the current request information includes: In response to current request information input to the graphical user interface, controlling the real-time search engine to retrieve at least one real-time news event corresponding to the current request information; Extracting corresponding key information from each of the real-time news events, and inputting the key information into the multimodal big model, and determining summary information corresponding to each of the real-time news events; wherein the multimodal big model includes a first preset template for extracting the summary information, and the first preset template is determined according to the format requirements and attribute requirements of the summary information; Determining the authenticity of the real-time news event corresponding to each of the summary information according to the verification rules and verification weights corresponding to each news element of the summary information in the third preset template; Determining a target real-time news event corresponding to the current request information from the real-time news events according to the authenticity of the real-time news event and the summary information; The current request information and summary information of the target real-time news event are input into the multimodal macro model, and a response to the current request information is output under the action of the second preset template.

6. The method according to claim 1, characterized in that The step of determining the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model includes: Determine, according to the preset format requirements and preset attribute requirements of the current summary information, a first preset template for each real-time news event corresponding to the current request information; wherein the preset attribute requirements include one or more of the following: news elements of the real-time news event required to be displayed, the number of words of the summary required to be displayed, and the news type of the real-time news event required to be displayed; Extracting corresponding key information from each of the real-time news events and inputting it into a multimodal large model; Extracting corresponding first key information from the key information of each real-time news event according to the preset attribute requirements; Each of the first key information is processed according to the preset format requirements to obtain summary information of each of the real-time news events; wherein the preset format requirements include a subject, a predicate, an object and a qualifier, and the qualifier is located before the subject and the object.

7. The method according to claim 1, characterized in that The step of inputting the multimodal macro model according to the first response, the second response and the historical response to determine the response to the current request information in the multi-round dialogue includes: Based on the relevance of the current request information with the first response, the second response and the historical response, respectively assigning a weight ratio to the first response, the second response and the historical response; According to the matching degrees and the weight ratios of the first response, the second response and the historical response with the current request information respectively, the responses to the current request information in the multiple rounds of dialogue are determined and output.

8. A multi-round dialogue processing device for real-time news based on a large model, characterized in that: include: A judgment module, in response to current request information for real-time news events, judges whether the current request information constitutes a multi-round dialogue with the previous request information; The first response module controls the multimodal large model to output a response to the current request information if no; The first determination module, if yes, determines the current summary information of the real-time news event corresponding to the current request information based on the first preset template in the multimodal large model; A filling module fills the current summary information, current request information, historical summary information, historical request information and historical responses in the multi-round dialogue into a second preset template of the multimodal large model; A second determination module determines, based on a second preset template of the multimodal large model, a first response of each current summary information in the multiple rounds of dialogue under the action of each historical request information, and a second response of each historical summary information in the multiple rounds of dialogue under the action of each current request information; a second response module, inputting the multimodal large model according to the first response, the second response and the historical response, and determining a response to the current request information in the multi-round dialogue; The filling module is further used to obtain each historical request information in the multiple rounds of dialogue, the historical summary information of each real-time news event corresponding to each historical request information, and the historical response corresponding to each historical request information; configure the field group in the second preset template of the multimodal large model according to the preset length and preset content type; the field group includes a first field group and a second field group; Extract target information from each historical request information, the historical summary information of each real-time news event corresponding to each historical request information and the historical response corresponding to each historical request information, and each current request information and the current summary information of each real-time news event corresponding to each current request information; wherein the target information meets the requirements of the preset content type and the requirements of the preset length; automatically fill the target information of each historical request information and the target information of each current summary information into the first field group, and automatically fill the target information of each current request information and the target information of each historical summary information into the second field group; The second determination module is also used to traverse the current summary information corresponding to each historical request information in the first field group, and generate a first response for each current summary information under the action of the historical request information; traverse the historical summary information corresponding to each current request information in the second field group, and generate a second response for each historical summary information under the action of the current request information.

Images